The AmneziaWG sidecar failed at startup with "Physical interface for
Amnezia UDP transport not found". Three compounding causes:
- physical_network() spawned PowerShell and filtered on HardwareInterface,
which wrongly excludes Hyper-V / WSL / Docker vEthernet uplinks, so every
AWG key failed identically on such hosts; a transient miss aborted the
whole startup with a hard OSError.
- A successful tunnel changes the outbound IP (e.g. 10.9.0.49), which the
network monitor misread as a real network change and reconnected on,
re-running the resolve while the tunnel held the default route — a
self-inflicted flap.
Replace the PowerShell probe with a pure-WinAPI resolver: extend
win_netinfo (GetAdaptersAddresses gateways, DNS, OperStatus, IfType,
Ipv4Metric) and pick the up adapter that owns an IPv4 gateway and a
routable address, lowest Ipv4Metric. Requiring a gateway naturally
excludes the WireGuard/AWG TUN (it has none), so the relay never binds
back onto its own tunnel; dropping HardwareInterface fixes vEthernet
uplinks. The owned Go core still hard-requires a non-zero interface index
(IP_UNICAST_IF, no default-bind fallback), so a transient miss is retried
instead of aborting. Ignore network-change events whose address is the
active tunnel's own, breaking the reconnect loop.
Extract the shared sidecar seam — the loopback SOCKS relay must accept a
TCP connection before the sing-box front dials it — into
engines/sidecar.wait_for_loopback_relay(). Hysteria delegates to it (no
behavior change); Amnezia adopts it as a confirming gate after
relay_ready, so both cores share one readiness contract. The divergent
domain models (config handoff, handshake detection, failure taxonomy,
recovery) intentionally stay per-engine.
Verified on Linux: 890 unit tests pass (offscreen). Tunnel-comes-up and
Hysteria startup still need on-device confirmation on Windows.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HAjvZYzPW2yTtToKJXGdbS
loop-uh
7acad15dcd